This is a (reproduced) implementation of the method described in this paper. Note that this implementation is not currently capable of perfectly reproducing the results published in the paper, though the results are close; the reason for this is unknown. If you would like to contribute, please see the "contributions" section in this README.
Defaults: cosine h(x), denseNet, cifar10, imagenet crop, 300 epochs, cross-entropy loss
Main two files are cal.py and deconfnet.py
To train, simply call
python cal.py
If you would like to contribute to the project, feel free to open issues and make pull requests.